Optimizing Nonlinear Control Allocation
Tor A. Johansen 1
Abstract
Control allocation is commonly utilized in over-actuated me-
chanical systems in order to optimally generate a requested
generalized force using a redundant set of actuators. Using a
control-Lyapunov approach, we develop an optimizing con-
trol allocation algorithm in the form of a dynamic update law,
for a general class of nonlinear systems. The asymptotically
optimal control allocation in interaction with an exponen-
tially stable trajectory-tracking controller guarantees uniform
boundedness and uniform global exponential convergence.
1 Introduction
Consider the nonlinear system
ẋ = f (t, x, τ ) (1)
τ = h(t, x, u) (2)
where t ≥ 0 is time, x ∈ Rn is the state vector, u ∈ Rr
is the control input vector, and τ ∈ Rp is a vector of virtual
controls, typically moments and forces in mechanical sys-
tems. During control design, the virtual control τ is treated
as an available input, although it can only be manipulated in-
directly via the input u through τ = h(t, x, u). Mapping the
requested τ to an input u is the control allocation task. As-
sume given a virtual control τc , in terms of a state feedback
law
τc = k(t, x) (3)
that uniformly exponentially stabilizes the origin of the sys-
tem (1) with perfect control allocation, i.e. τ = τc . The state
x typically represents the tracking error relative to a time-
varying reference trajectory, possibly also including an expo-
nentially stable observer error. The basic control allocation
problem is then to solve the system of nonlinear algebraic
equations (2) with respect to the control vector u subject to
τ = τc . Since we consider fully- or over-actuated problems
(p ≥ r), this does not in general define a unique u and one
usually introduces an instantaneous cost function, J(t, x, u).
The cost function may incorporate power consumption or
cost of raw materials, for example, and we assume actuator
limitations and operational constraints are embedded into J
as penalty or barrier functions. The control allocation prob-
lem is then formulated in terms of solving the following non-
linear static minimization problem:
min J(t, x, u) subject to τc − h(t, x, u) = 0 (4)
u ||f (t, x1 , τ1 ) − f (t, x2 , τ2 )|| ≤ Lx ||x1 − x2 || + Lτ ||τ1 − τ2 ||
Optimizing solutions have been derived for certain classes of
over-actuated systems, such as aircraft, marine vessels, and
machines [1, 2, 3, 4, 5, 6, 7]. They all treat the control allo-
cation problem as a static (or quasi-dynamic) problem that is
solved independently of the dynamic control problem, gen-
erally considering linear models τ = Gu, with G ∈ Rp×r .
The main advantage of this is modularity through its hierar-
chical structure. In the present paper we take a Lyapunov-
based design approach, and consider general nonlinear mod-
els. Essentially, we specify a control Lyapunov function and
1 Department of Engineering Cybernetics, Norwegian University of
Science and Technology, N-7491 Trondheim, Norway. e-mail:
tor.arne.johansen@itk.ntnu.no.
derive an exponentially convergent dynamic update law for
u (similar to a gradient/Newton-like optimization or adap-
tive control law [8]) such that the control allocation problem
is solved dynamically. The main contribution is a theoretical
result that shows that it is not necessary to solve the optimiza-
tion problem (4) exactly at each time instant (or sampling
instant in a discrete-time implementation). It is shown that
convergence and asymptotic optimality of the dynamic con-
trol allocation in combination with an exponentially stable
trajectory-tracking nonlinear controller guarantees uniform
boundedness and uniform global exponential convergence of
the system. This is related to the concept of asymptotic op-
timality [9, 10] and sub-optimal control [11]. One advan-
tage of this approach is computational efficiency, since the
optimizing control allocation algorithm is implemented ex-
plicitly as a dynamic nonlinear controller with r + p states
to update. Solving (4) explicitly at each sampling instant
requires a computationally more expensive numerical solu-
tion of a nonlinear program to guarantee optimality, [12],
although the present results indicate that the computational
complexity can be safety reduced by for example early ter-
mination of the iterative numerical optimization.
2 Lyapunov design
Assumption 1. The virtual controller (3) makes the origin
uniformly globally exponentially stable, i.e. there exists a
differentiable function V0 : [0, ∞) × Rn → R and positive
constants c1 , c2 , c3 and c4 such that
c1 ||x||2 ≤ V0 (t, x) ≤ c2 ||x||2 (5)
∂V0 ∂V0T
(t, x) + (t, x)f (t, x, k(t, x)) ≤ −c3 ||x||2 (6)
∂t ∂x ¯¯ ¯¯
¯¯ ∂V0 ¯¯
¯¯ ¯¯
¯¯ ∂x (t, x)¯¯ ≤ c4 ||x|| (7)
Assumption 2. There exists a constant % > 0 such that
∂h ∂hT
(t, x, u) (t, x, u) ≥ %Ip (8)
∂u ∂u
Assumption 3. The function f is differentiable and satis-
fies f (t, 0, 0) = 0. Moreover, it is globally Lipschitz, uni-
formly in t, i.e. there exist constants Lx and Lτ such that
for all x1 , x2 ∈ Rn , τ1 , τ2 ∈ Rp and t ≥ 0. The function
h is twice differentiable and globally Lipschitz, uniformly in
t, with h(t, 0, 0) = 0 and Lipschitz constant Lh in x and
u. Finally, we require that k is differentiable and Lipschitz,
uniformly in t, with k(t, 0) = 0.
Assumption 4. The cost function J is twice differentiable.
The optimization problem (4) is reformulated by introducing
a vector of Lagrange multipliers λ ∈ Rp and the Lagrangian
`(u, λ, x, t) = J(t, x, u) + (τc − h(t, x, u))T λ (9)
Local minima of (4) satisfy the first order optimality condi-
tions for `, and we define the limiting optimal set E ∗ accord-
ingly: Lemma 1 Suppose assumptions 2 – 5 hold. Then α = 0 and
© ¯ ∂`
β = 0 is equivalent to ∂u ∂`
= 0 and ∂λ = 0.
E ∗
= (x, u, λ) ∈ Rn+r+p ¯ x = 0,
¾
∂` ∂` Proof. Lemma 16.1 in [13] proves that H is bounded away
lim (u, λ, x, t) = 0, lim (u, λ, x, t) = 0
t→∞ ∂u t→∞ ∂λ from singularity due to Assumptions 2 and 5. ¤
For simplicity, it is assumed that the limit exists. The fol- Eq. (11) can be rewritten in the compact form
lowing control Lyapunov function is designed to attract the
total state (x, u, λ) to E ∗ (notice that u and λ are yet un- ∂V0
V̇ = σ (t, x)
specified, but will be states in the dynamic control allocation ∂t
algorithm): ∂V T
µ ¶ +σ 0 (t, x)f (t, x, k(t, x) + h(t, x, u) − k(t, x))
1 ∂`T ∂` ∂`T ∂` ∂x
V (t, x, u, λ) = σV0 (t, x) + + (10)
2 ∂u ∂u ∂λ ∂λ +αT u̇ + β T λ̇ + δ (15)
where σ > 0 is a constant that will be specified later. The The αT u̇ term in (15) is made negative definite by the first
arguments of ` are implicit in (10) to simplify the notation. term of the dynamic update law
The time-derivative of V along trajectories of (1) satisfies
∂V0 u̇ = −Γα + ζ (16)
V̇ = σ (t, x)
∂t
∂V T With Γ = ΓT > 0. Similarly, the β T λ̇ term in (15) is made
+σ 0 (t, x)f (t, x, k(t, x) + h(t, x, u) − k(t, x)) negative by the first term of the dynamic update law
∂x
µ T 2 ¶
∂` ∂ ` ∂`T ∂ 2 ` λ̇ = −W β + φ (17)
+ + u̇
∂u ∂u2 ∂λ ∂u∂λ
µ T 2 ¶ µ T 2 ¶ with W = W T > 0. The last (indefinite) term in (15) is
∂` ∂ ` ∂` ∂ ` ∂`T ∂ 2 `
+ λ̇ + + cancelled if the vector signals ζ(t) ∈ Rr and φ(t) ∈ Rp can
∂u ∂λ∂u ∂u ∂x∂u ∂λ ∂x∂λ be chosen such that the following scalar algebraic equation
∂`T ∂ 2 ` ∂`T ∂ 2 ` holds:
·f (t, x, h(t, x, u)) + + (11)
∂λ ∂t∂λ ∂u ∂t∂u αT ζ + β T φ + δ = 0 (18)
where
We will return to this issue shortly, and show that we can
∂` ∂J ∂hT always find signals ζ and φ such that this equation holds for
(u, λ, x, t) = (t, x, u) − (t, x, u)λ (12)
∂u ∂u ∂u all t ≥ 0. Using theorem 2.4.7 in [14], we get from (15),
∂` (16), and (17) with the algebraic constraint (18):
(u, λ, x, t) = τc − h(t, x, u) (13)
∂λ µ ¶
∂V0 ∂V0T
Define for notational convenience V̇ = σ (t, x) + (t, x)f (t, x, k(t, x))
∂t ∂x
∂ 2 ` ∂` ∂hT ∂`
α = − −αT Γα − β T W β
∂u2 ∂u ∂u ∂λ
∂h ∂` ∂V T
β = − +σ 0 (t, x)R(t, x, u, τc )(h(t, x, u) − τc )(19)
∂u ∂u ∂x
µ T 2 ¶
∂` ∂ ` ∂`T ∂h with
δ = − f (t, x, h(t, x, u))
∂u ∂x∂u ∂λ ∂x Z
µ ¶ 1
∂f
∂`T ∂h ∂`T ∂ 2 ` R(t, x, u, τc ) = (t, x, sτc + (1 − s)h(t, x, u))ds
+ τ̇c − + 0 ∂τ
∂λ ∂t ∂u ∂t∂u
(20)
and observe that
∂2` ∂hT ∂2` ∂h ∂τc The following global convergence result shows that the last
=− , =− + , (indefinite) term in (19) is dominated by the other (negative)
∂u∂λ ∂u ∂x∂λ ∂x ∂x terms.
p
∂2` ∂ 2 J X ∂ 2 hi
= − λi 2
∂u2 ∂u2 i=1
∂u Proposition 1 Consider the system (1), (2), (3), (16) and
(17) with ζ(t) and φ(t) satisfying (18). If Assumptions 1
Hence, the definitions of α and β can be written - 5 hold, then ||x(t)||, ||τ (t)|| and ||τc (t)|| are uniformly
³ ´ µ ∂` ¶ µ ∂2` T ¶ bounded, and (x(t), u(t), λ(t)) → E ∗ as t → ∞ with ex-
α − ∂h ponential convergence rate, for any initial conditions x(0) ∈
β =H , with H =
∂u ∂u 2 ∂u (14)
∂` ∂h
∂λ − ∂u 0 Rn , λ(0) ∈ Rp , and u(0) ∈ Rr .

Assumption 5. There exist constants κ2 > κ1 > 0 such that Proof. First, we show that V is radially unbounded. Consider
∂2`
κ1 Ir ≤ ∂u 2 ≤ κ2 Ir . any x0 ∈ Rn , and some optimal u0 ∈ Rr and λ0 ∈ Rp such
that (x0 , u0 , λ0 ) ∈ E ∗ . Using theorem 2.4.7 in [14] it is this time-varying scalar algebraic equation we solve a least-
straightforward to show that squares problem subject to (18). This leads to the Lagrangian
¯ ¯ 1¡ T ¢ ¡ ¢
∂` ∂ 2 ` ¯¯ ∂h T ¯¯ L(ζ, φ, ν) =
2
ζ ζ + φT φ + ν αT ζ + β T φ + δ (28)
= (u − u0 ) − ¯ (λ − λ0 ) (21)
∂u ∂u2 ¯0 ∂u ¯ where ν ∈ R is a Lagrange multiplier. First order optimality
0
¯ conditions leads to ζ and φ being given by the solution to the
∂` ∂h ¯¯
= − (u − u0 ) (22) following time-varying linear system of equations:
∂λ ∂u ¯0 Ã !Ã ! Ã !
Ir 0 α ζ 0
¯ R 1 ∂2` ¯ 0 Ip β
∂2` ¯ ∂h ¯ φ = 0 (29)
where ∂u 2¯ = ∂u2 (su0 +(1−s)u, λ, x, t)ds and ∂u 0 =
0 0 αT β T 0 ν −δ
R 1 ∂h
0 ∂u
(su0 + (1 − s)u, λ, x, t)ds. Hence,
Lemma 2 Suppose Assumptions 2, 3 and 5 hold. Then (29)
à ¯ ¯ ! always has a unique solution for ζ and φ.
³ ´ ∂2` ¯ ∂hT ¯
∂`T ∂` ∂`T ∂` u − u0 T ∂u 2 ¯ − ∂u ¯
+ = λ−λ ¯0 0 Proof. Assuming α 6= 0 or β 6= 0, the solution is indeed
∂u ∂u ∂λ ∂λ 0 ∂h ¯
− ∂u 0 0 unique, see Lemma 16.1 in [13]. On the other hand, if α = 0
à ¯ ¯ ! and β = 0 it is evident from the definition of δ that δ = 0, due
∂2` ¯ T ¯
³ ´
∂u 2¯ − ∂h∂u ¯ u − u0 to Lemma 1. The last equation in (29) then becomes trivial
· ¯0 0
λ − λ0
∂h ¯
− ∂u 0 and φ = 0, ζ = 0 defines the solution (notice that ν is not
0 uniquely defined in this case). ¤
Using Assumptions 2 and 5 it follows immediately that there Proposition 2 Under the assumptions of Proposition 1, λ(t)
exists a constant k0 > 0 such that is uniformly bounded.
∂`T ∂` ∂`T ∂` ¡ ¢ Proof. Substituting into (17) gives
+ ≥ k0 ||u − u0 ||2 + ||λ − λ0 ||2 (23)
∂u ∂u ∂λ ∂λ ∂h ∂hT
λ̇ = −W λ+χ (30)
Since x0 is arbitrary we conclude from Assumption 1 that ∂u ∂u
V is radially unbounded in (u, λ, x). Next, from (19) and ∂h ∂J
Assumption 1 it follows that with χ = W ∂u ∂u + φ. Consider the Lyapunov-like function
1 T
V(λ) = 2 λ λ. Its time-derivative is
V̇ ≤ −c3 σ||x||2 − λmin (W )||α||2 − λmin (Γ)||β||2 ∂h ∂hT
¯¯ ¯¯ V̇ = −λT W λ + λT χ
¯¯ ∂` ¯¯ (31)
+2σLτ ¯¯¯¯ ¯¯¯¯ c4 ||x|| (24) ∂u ∂u
∂λ ≤ −%λmin (W )||λ||2 + ||λ|| · ||χ|| (32)
From the definitions of α and β together with Assumptions Uniform boundedness of λ(t) follows because %λmin (W ) >
2, 3 and 5 one can derive 0 and Proposition 1 and Lemma 2 implies that ||χ|| is uni-
¯¯ ¯¯ formly bounded. ¤
¯¯ ∂` ¯¯ 2
¯¯ ¯¯ ≤ Lh ||α|| + κ2 Lh ||β|| (25) Remark 1. Notice that the matrices Γ > 0 and W > 0
¯¯ ∂λ ¯¯ % %2 may be chosen as time-varying, without changing any of the
theoretical properties, provided they are bounded away from
³ ´ zero. Newton-like methods can therefore be implemented by
L L2 c κ
Let M = max Lτ L%h c4 , τ %h2 4 2 . Using Young’s in- taking
equality 2ab ≤ a2 /µ + b2 µ for µ > 0, (24) and (25) lead µ ¶
u̇ ¡ ¢−1 ³ α ´ ³ ζ ´
to = −γ HT H + εIr+p β + φ (33)
λ̇
V̇ ≤ −σ(c3 − M µ)||x||2 − (λmin (W ) − σM/µ)||α||2 where γ > 0 and ε ≥ 0 are time-varying parameters. In a
−(λmin (Γ) − σM/µ)||β||2 (26) discrete-time implementation, γ may be chosen using a line
search to guarantee descent between each sampling instant
Notice that µ > 0 and σ > 0 are arbitrary constants. First, (in terms of a merit function), [13].
choose µ > 0 such that c3 > M µ. Next, choose σ > 0 such Remark 2. The terms involved in the algebraic constraint
that λmin (W ) − σM/µ > 0 and λmin (Γ) − σM/µ > 0. Be- (18) arise because the optimal solution u to (4) is time-
cause H is bounded away from singularity (due to Assump- varying. The terms ζ and φ provides a feedforward-like com-
tions 2 and 5), there exist constants k1 , k2 , k3 , k4 > 0 such pensation in the update laws for u and λ, seeking to maintain
that the time-varying optimum.
¯¯ ¯¯2 ¯¯ ¯¯2 Remark 3. Although λ(0) can be chosen arbitrarily, one can
¯¯ ∂` ¯¯ ¯¯ ∂` ¯¯
2 ¯ ¯
V̇ ≤ −k1 ||x|| − k2 ¯¯ ¯¯ − k3 ¯¯¯¯ ¯¯¯¯ ≤ −k4 V (27)
¯ ¯ reduce transients by λ(0) = arg minλ V (0, x(0), u(0), λ).
∂u ∂λ Remark 4. If E ∗ contains a unique minimum (for example
under some strict convexity assumption of J and additional
Uniform boundedness and exponential convergence follow assumptions on h), or the dynamics are time-invariant, one
directly. ¤ may extend the convergence result in a standard manner to
Consider the issue of solving (18) with respect to ζ ∈ Rr global exponential stability. Robustness is then an inherent
and φ ∈ Rp . To achieve a well-defined unique solution to property.
 3 Objectives and constraints It is straightforward to verify that the system matrix in (36)
Usually, one wants to introduce a cost function J(t, x, u) that is Hurwitz. Notice that ζ = 0 and φ = 0 must hold at the
captures multiple objectives such as minimizing power con- equilibrium point (origin). If x is considered as a constant
sumption, satisfying input constraints, and avoiding singular- input in (36), it is easily verified that
ities. For example, power consumption can in some cases be
approximated with the following term [3] u = H −1 GT (GH −1 GT )−1 (−Kx) (37)
1 T λ = −(GH −1 GT )−1 (−Kx) (38)
J1 (u) = u Hu
2 defines the equilibrium point for (36). The solution (37)
Input constraints on the form c(u) ≤ 0 can be added to the coincides with the conventional generalized inverse solu-
optimization problem through a barrier function of the form tion, e.g. [1], (or the Moore-Penrose pseudo-inverse G+ =
X GT (GGT )−1 if H = Ir ). Using standard singular perturba-
J2 (u) = −w2 log(−ci (u)) (34) tion arguments (assuming the (u, λ)-dynamics are faster than
i the x-dynamics), it follows that the optimal solution (37) is
asymptotically attained and that the suggested approach only
with w2 > 0, [13]. J2 will not be defined outside the admis- differs from from the generalized inverse solution by a fast
sible region, so Proposition 1 reduces to a local convergence transient term due to the initial conditions u(0) and λ(0) not
result. necessarily satisfying the optimality conditions.
Since the control allocation algorithm explicitly computes u̇,
input rate constraints can be enforced by reducing the gain 4.2 Low-speed manoeuvering of over-actuated ship
Γ(t) sufficiently. Again, Proposition 1 may be reduced to a This example is based on [15]. Consider a ship equipped
local convergence result because (ζ, φ) (which are not influ- with two rudder/propeller pairs at the stern, and one tunnel
enced by Γ) may require unacceptable high input rates, and thruster at the bow. The nonlinear equations of motion in the
in addition one can in general not find a strictly positive lower horizontal plane are
bound on inf t λmin (Γ(t)) ≥ 0.
Singular effector configurations must usually be avoided be- η̇ = R(ψ)ν (39)
cause they may lead to temporary loss of full controllability, M ν̇ + Dν = τ +d (40)
[12]. This can be implemented by adding the following non-
convex term to the criterion where η ∈ R3 contains the (x, y)-position and heading ψ
µ ¶
∂h ∂hT in an Earth-fixed coordinate frame, and ν ∈ R3 the corre-
J3 (t, x, u) = −w3 log det (t, x, u) (t, x, u) + ²Ip sponding velocity components in a vessel-fixed coordinate
∂u ∂u frame. M is an inertia matrix, D is a linear damping matrix,
with w3 > 0, and ² > 0. Notice that a finite value of τ ∈ R3 contains surge and sway forces, as well as the yaw
J3 (t, x, u) implies that Assumption 2 is always satisfied if momentum, and d ∈ R3 is a vector of slowly time-varying
² = 0. Hence, using this term in the criterion makes the con- disturbances. R(ψ) is the rotation matrix from the vessel-
trol allocation algorithm avoid values of u where the effector fixed frame to the Earth-fixed frame. It is shown in [15] that
configuration is singular, such that assumption 2 is effectively the following controller is globally exponentially stabilizing
enforced. under a reasonable assumption on ψ:
4 Examples τc = −KI RT (ψ)ξ − KP RT (ψ)(η − η ∗ ) − KD ν(41)
4.1 Linear state feedback control with linear effector ξ˙ = η − η∗ (42)
model
The example is intended to illustrate that the suggested ap- The model and controller parameters in the simulation study
proach is a natural nonlinear extension to the generalized in- are chosen in accordance with the scale model ship studied
verse solution. We consider the over-actuated linear system in [15]. At low speed, the surge force X (longitudinal) and
ẋ = Ax + Bτ, τ = Gu (35) sway force Y (lateral) produced by a propeller/rudder pair is
given by [15]
where the controller τc = −Kx is stabilizing such that ½
A−BK is Hurwitz. We assume G has full rank, such that the kT p ω 2 , ω≥0
control allocation problem is non-singular and assumption 2 T = (43)
kT n |ω| ω, ω < 0
holds. If G does not have full rank, then the system (A, BG)
might not be controllable even if (A, B) is controllable.
Hence, this assumption is not restrictive. The cost func- ½
tion is the standard quadratic form J(u) = 12 uT Hu, with T (1 + kLn ω)(kLδ1 δ + kLδ2 |δ| δ), ω ≥ 0
Lr = 0, ω<0 (44)
∂` ∂`
H = H T > 0, and it follows that ∂u = Hu − GT λ, ∂λ = ½
τc − Gu. This leads to α = HHu − HGT λ − GT (τc − Gu), Dr = T (1 + kDn ω)(kDδ1 |δ| + kDδ2 δ 2 ), ω ≥ 0 (45)
β = −G(Hu − GT λ), and δ = −(τc − Gu)T Kx. The 0, ω<0
control allocation algorithm is
µ ¶ µ ¶³ T is the nominal thrust, Lr is the rudder lift force, Dr is the
u̇ ´ rudder drag force, ω is the propeller angular velocity, and δ is
−Γ(HH + GT G) ΓHGT u
= T λ the rudder angle. The surge and sway forces of each thruster
λ̇ W GH −W GG
µ ¶ are given as follows:
T
+ −ΓG Kx + ζ (36) X = T − Dr , Y = Lr
φ (46)
 1.5 0.04

1 0.02
η1

ν1
0.5 0

0 −0.02

−0.5 −0.04
0 100 200 300 400 500 600 0 100 200 300 400 500 600

1.5 0.04

1 0.02
η2

ν2
0.5 0

0 −0.02

−0.5 −0.04
0 100 200 300 400 500 600 0 100 200 300 400 500 600

0.3 0.03

0.2 0.02
η3

ν3
0.1 0.01

0 0

−0.1 −0.01
0 100 200 300 400 500 600 0 100 200 300 400 500 600
t t
Figure 1: Simulation results - solid lines are positions while Figure 2: Simulation results - velocities.
dashed lines are reference.

For the bow tunnel thruster, the same model can be used with
rudder angle δ ≡ 0. Let the stern propeller/rudder pairs have
index 1 and 2, and the bow tunnel thruster have index 3. Thus
the virtual controls τ = (τ1 , τ2 , τ3 )T defined by
τ1 = X1 + X2 (47)
τ2 = Y1 + Y2 + T3 (48)
τ3 = −`1,y X1 + `1,x Y1 − `2,y X2 + `2,x Y2 + `3,x T3(49)
are related to the control signals u = (ω1 , ω2 , ω3 , δ1 , δ2 )T ∈
R5 via the nonlinear model equations (43)- (46). The mo-
ment arms in (49) are defined by the location of the propul-
sion devices. Constraints |ωi | ≤ 18 Hz, and |δi | ≤ 0.61 rad
(35 deg) are implemented as a barrier function J2 on the form
(34) with w2 = 0.01. In addition, the cost function measures
relative power consumption and use of rudders
3
X 2
X [1] W. C. Durham, “Constrained control allocation,” J.
J(u) = ki |ωi |ωi2 + qj δj2
i=1 j=1
with k1 = k2 = 0.01, k3 = 0.02, and q1 = q2 = 500.
In the simulation example, we use (33) for optimization, with
γ = 1, ² = 10−9 . The simulation results are presented in Fig-
ures 1-5, with a constant disturbance d. Except for the initial
transient (due to an arbitrary choice of u(0) and λ(0)) and a
short period around t ≈ 210 where ω3 saturates, we observe
from Figure 3 that the attained generalized forces τc are very
close to the commanded generalized forces τ . The control al-
location problems is a non-convex optimization problem due
to asymmetry in rudder lift with positive and negative surge
thrust, cf. (44) and [15]. Hence, the control allocation al-
gorithm only seeks a locally optimal solution. One benefit
of the dynamic control allocation algorithm compared to a
(quasi-)static approach is that one avoids chattering due to
the fact that globally optimal control allocation is in fact dis-
continuous as a function of the requested generalized forces
[15, 16].
5 Concluding remarks
Taking a control-Lyapunov design approach, an optimizing
nonlinear control allocation algorithm is derived. The al-
gorithm leads to asymptotic optimality, thus relaxing the
computational complexity considerably compared to a di-
rect nonlinear programming approach. At the same time,
we guarantee global exponential convergence of the overall
system comprising an uniformly globally exponentially sta-
ble trajectory-tracking nonlinear controller together with the
control allocation algorithm. It is also interesting to observe
that the method leads to feedforward-like terms that takes
into account the fact that the optimum is time-varying.
Acknowledgements
The author is grateful to Thor I. Fossen and Roger Skjetne
for insightful comments and suggestions for improvements.
References
+ J2 (u) (50) Guidance, Control and Dynamics, vol. 16, pp. 717–725,
1993.
[2] D. Enns, “Control allocation approaches,” in Proc.
AIAA Guidance, Navigation and Control Conference and Ex-
hibit, Boston MA, 1998, pp. 98–108.
[3] O. J. Sørdalen, “Optimal thrust allocation for marine
vessels,” Control Engineering Practice, vol. 5, pp. 1223–
1231, 1997.
[4] S. P. Berge and T. I. Fossen, “Robust control allocation
of overactuated ships; experiments with a model ship,” in
Preprints IFAC Conference on Maneuvering and Control of
Marine Craft, Brijuni, Croatia, 1997.
[5] W. Jackson, M. P. J. Fromherz, D. K. Biegelsen,
J. Reisch, and D. Goldberg, “Constrained optimization based
control of real time large-scale systems: Airjet object move-
ment system,” in Proc. IEEE Conf. Decision and Control,
Orlando, 2001.
[6] M. Bodson, “Evaluation of optimization methods for
 0.2
0
0.1
−5
τ1

ω1
−10
−0.1
−15
−0.2
0 100 200 300 400 500 600 −20
0 100 200 300 400 500 600
0.4
20
0.2
10
τ2

ω2
−0.2 0

−0.4
0 100 200 300 400 500 600 −10
0 100 200 300 400 500 600
0
20

−0.02
10
τ3

ω3
−0.04
0

−0.06
0 100 200 300 400 500 600 −10
0 100 200 300 400 500 600
t
t
Figure 3: Simulation results - solid lines are attained general- 0

ized forces and dashed lines are generalized forces com- −5

manded by the controller.
δ1 (deg) −10

−15

−20
control allocation,” J. Guidance, Control and Dynamics, vol. 0 100 200 300 400 500 600

25, pp. 703–711, 2002.

30

[7] O. Härkegård, “Efficient active set algorithms for solv- 20

ing constrained least squares problems in aircraft control al-

δ2 (deg)

10
location,” in Proc. IEEE Conf. Decision and Control, Las 0
Vegas NV, 2002.
−10

[8] M. Krstic, I. Kanellakopoulos, and P. Kokotovic, Non- −20

linear and Adaptive Control Design, Wiley and Sons, 1995. 0 100 200 300
t
400 500 600

[9] M. Cannon and B. Kouvaritakis, “Efficient constrained Figure 4: Simulation results - control signals computed by the con-
model predictive control with asymptotic optimality,” SIAM trol allocation algorithm.
J. Optimization and Control, vol. 41, pp. 60–82, 2002.
[10] T. A. Johansen and D. Sbárbaro, “Optimizing control
of over-actuated linear systems with nonlinear output maps
via control Lyapunov functions,” in European Control Con- 50

ference, Cambridge, UK, 2003. 0

[11] P. O. M. Scokaert, D. Q. Mayne, and J. B. Rawlings,

λ1

−50
“Suboptimal model predictive control (feasibility implies sta-
−100
bility),” IEEE Trans. Automatic Control, vol. 44, pp. 648–
654, 1999. −150
0 100 200 300 400 500 600
[12] T. A. Johansen, T. I. Fossen, and S. P. Berge, “Con- 1500
strained nonlinear control allocation with singularity avoid-
1000
ance using sequential quadratic programming,” IEEE Trans.
λ2

Control Systems Technology, accepted, 2003. 500

[13] J. Nocedal and S. J. Wright, Numerical Optimization, 0

Springer-Verlag, New York, 1999. −500

0 100 200 300 400 500 600
[14] R. Abrahamson, J. E. Marsden, and T. Ratiu, Man-
1500
ifolds, Tensor Analysis and Applications (2nd Edition),
Springer-Verlag, New York, 1988. 1000
λ3

[15] K.-P. Lindegaard and T. I. Fossen, “Fuel efficient con- 500

trol allocation for surface vessels with active rudder usage: 0

Experiments with a model ship,” IEEE Trans. Control Sys-
−500
tems Technology, vol. 11, pp. 850–862, 2003. 0 100 200 300 400 500 600

[16] T. A. Johansen, T. P. Fuglseth, P. Tøndel, and T. I. Fos- t

sen, “Optimal constrained control allocation in marine sur- Figure 5: Simulation results - Lagrange multipliers.
face vessels with rudders,” in IFAC Conf. Manoeuvring and
Control of Marine Craft, Girona, 2003.